If you build or deploy AI systems that assist courts or judges in researching facts and law, applying law to concrete cases, or that function in a similar way in alternative dispute resolution — EU AI Act Annex III Point 8 classifies your system as high-risk and triggers full Title III obligations before the August 2026 general application deadline. Administration of justice AI operates at the intersection of the most constitutionally protected fundamental right in the EU Charter: Art.47 (right to an effective remedy and fair trial), Art.48 (presumption of innocence), and Art.49 (legality and proportionality of criminal offences and penalties). The Art.6 ECHR fair trial guarantee — which requires judicial decisions to be taken by an independent human judge with individual reasoning — creates a dual constitutional constraint alongside EU AI Act Art.14(4): no judicial AI system may automate the final act of adjudication, regardless of how sophisticated its legal research or fact interpretation capabilities are. The sovereignty gap is real: Harvey AI (San Francisco), Casetext (acquired by Thomson Reuters, US-entity), and Luminance (Cambridge, UK, with US institutional investors) all have CLOUD Act exposure on EU court case files, judgments under preparation, and attorney-client privileged communications if deployed in EU judicial contexts without sovereignty-compliant data governance.

What Annex III Point 8 Actually Covers

Annex III Point 8 applies to AI systems used in the administration of justice and democratic processes. The judicial AI subcategory covers AI systems intended to be used by a competent authority or on behalf of such authority to assist a court or a judge to research and interpret facts and the law and to apply the law to a concrete set of facts, or to be used in a similar way in alternative dispute resolution.

Four elements define the Point 8 scope:

"Competent authority" context: The AI must be intended for use by a court, judge, tribunal, arbitration panel, or administrative adjudication body — or deployed on behalf of such authority (e.g., a court-appointed technical assistant using AI tools in court-ordered expert assessments). Legal research AI used exclusively by private law firms representing clients does not trigger Point 8, even if it is technically capable of being used in a judicial context.

"Assist" not "decide": Point 8 does not require that the AI make judicial decisions — it covers AI that assists the judge in researching, interpreting, and applying law. This means legal research tools (case law retrieval, statute interpretation, precedent analysis), AI that identifies relevant facts from large document sets, and AI that generates draft legal reasoning for judicial review all qualify as high-risk if deployed in a court context.

"Facts and the law": The AI must assist with factual or legal analysis relevant to the case at hand. Generic scheduling software, case management workflow tools, or translation AI used in courtrooms are not Point 8 AI, because they do not assist with the substantive legal analysis. The threshold is whether the AI output feeds into the judge's legal reasoning about the merits of the case.

"Similar way in alternative dispute resolution": Arbitration, mediation, and online dispute resolution (ODR) AI that assist arbitrators or mediators in fact-finding, legal analysis, or applying agreed rules to disputed facts qualifies under Point 8 by analogy. The test is functional equivalence to court assistance — not the formal institutional setting.

The "Assist" Not "Decide" Distinction — Court Context vs. Law Firm Context

The most operationally significant scope question for AI providers building legal research tools is where the Point 8 boundary falls relative to tools used by law firms rather than courts.

The key insight for providers: The same AI tool — Harvey AI, Luminance, or a custom legal research LLM — may or may not be high-risk under Annex III Point 8 depending entirely on who deploys it and for what role. A contract sold to a law firm is not Point 8; the same product licensed directly to the court administration triggers the full Point 8 conformity assessment obligation on the provider.

Provider obligations triggered: Once an AI system is intended for use by competent authorities in a judicial context, the provider (Harvey AI, Luminance, the SaaS vendor) bears Art.16 provider obligations regardless of whether the court itself initiated the deployment or whether the provider actively markets to courts. "Intended to be used" is determined by the provider's reasonable foreseeability of use — marketing materials, court-specific contract templates, or known court deployments all establish intent.

Legal Research AI as Provider vs. Deployer: Harvey, Luminance, Casetext

Three major legal AI platforms have achieved significant EU market penetration as of 2026, each presenting distinct EU AI Act Point 8 compliance profiles.

Harvey AI (Harvey Industries Inc., San Francisco, California): Harvey AI is built on Claude and GPT-4 class models, marketed to law firms for legal research, contract analysis, and draft generation. Harvey has expanded into court-adjacent contexts in several EU jurisdictions through partnerships with court administration bodies and legal aid providers. US-entity status = full CLOUD Act exposure: if Harvey processes EU judicial case files, ongoing litigation strategies, or pre-judgment draft reasoning on US-located infrastructure, DOJ subpoena power applies to all that content under 18 U.S.C. §2703. Harvey's standard terms do not provide EU judicial sovereignty guarantees. Point 8 classification: HIGH-RISK if deployed for court use; NOT HIGH-RISK if law firm only.

Luminance (Luminance Technologies Ltd., Cambridge, UK): Luminance uses its own proprietary legal AI trained on legal datasets, deployed in due diligence, contract analysis, and compliance review. Post-Brexit UK entity with US institutional investors (General Atlantic, TCV) — US investor structure creates indirect GDPR Art.44+ transfer risk if investor access to data is contractually permitted. Luminance has actively pursued EU court and regulatory body contracts. UK AI regulation is currently less prescriptive than EU AI Act for judicial AI. When Luminance is deployed in EU arbitration proceedings or administrative tribunals, the EU AI Act Point 8 high-risk classification applies regardless of Luminance's UK domicile — Art.2(1)(c) applies EU AI Act obligations to systems placed on the EU market or affecting EU persons. Point 8 classification: HIGH-RISK for EU arbitration/court deployments; provider obligations apply.

Casetext / CoCounsel (Thomson Reuters, Toronto/New York): Casetext was acquired by Thomson Reuters in 2023 and rebranded as CoCounsel. Thomson Reuters = US-entity with CLOUD Act exposure. CoCounsel is positioned for attorney use, not court administration — reducing Point 8 risk. However, court-appointed referees, legal aid coordinators employed by court systems, and judicial assistants using CoCounsel to assist judges in case research create Point 8 deployment contexts. Point 8 classification: Context-dependent — HIGH-RISK if deployed in court-assistance role; provider obligations follow.

EU-Native Alternative — Legalese (Riga, Latvia): Legalese is an EU-incorporated legal AI platform trained specifically on EU law and Member State legal corpora. No CLOUD Act exposure (EU entity, EU-hosted). Direct alignment with EU AI Act Annex III Point 8 compliance architecture. Represents the sovereignty-compliant path for EU court administration bodies procuring legal research AI.

COMPAS Equivalents in European Courts — Recidivism AI and Bail/Sentencing Support

The COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) risk assessment instrument is the most studied example of algorithmic risk scoring in criminal justice. Northpointe/Equivant's COMPAS is not widely used in EU courts — but equivalent risk assessment instruments have been developed and deployed across EU Member States.

HART (Harm Assessment Risk Tool) — Durham Constabulary, England/Wales: HART predicts whether a defendant is at low, medium, or high risk of reoffending over a two-year period. Used to inform pre-charge bail decisions. The Bridges v. Chief Constable of South Wales (2020) and Osborn v. The Parole Board (2013) cases established that algorithmic risk tools used in UK criminal justice require human oversight and individual justification. While HART operates in a UK context, its design framework is the reference model for several EU police-judiciary pilot projects. EU AI Act Point 8(a) classification: HIGH-RISK — the tool assists judicial/prosecutorial authorities in applying risk law to individual facts.

SkillMap / Criminological Risk Assessment (Netherlands — NSCR/WODC): The Netherlands Forensic Institute and NSCR have developed risk assessment instruments used in pre-trial detention hearings and parole evaluations. Dutch courts use these as advisory tools — the judge makes the final decision. Under EU AI Act Point 8, these instruments are HIGH-RISK if intended for use by the court (as deployer) or provided to courts for that purpose (as provider obligation).

KTOR (Kritisch Terugval Onderzoek en Risicobeoordeling) — Belgium: Belgian forensic psychology assessment tool used in criminal proceedings for offender risk profiling. Deployed by forensic experts acting on behalf of courts (court-appointed assessors) = "on behalf of competent authority" — triggering Point 8 high-risk classification.

BGH Case Law on Algorithmic Decisions: The German Bundesgerichtshof established in BGH I ZR 225/21 ("Scoring" case, 2023) that automated scoring systems used to make decisions affecting individuals require transparency, human oversight, and the ability to challenge the algorithm's output. The BGH reasoning — while originating in a civil law (SCHUFA credit scoring) context — establishes that German courts will apply similar principles to judicial AI tools used in criminal or civil proceedings.

The Loomis Precedent and Its EU Limits: The Wisconsin Supreme Court in State v. Loomis (2016) held that using COMPAS in sentencing did not violate due process because the judge was not required to follow the algorithm's recommendation. This US precedent does not translate to EU contexts: Art.6 ECHR requires not just that the judge retain formal authority, but that judicial reasoning be individually articulated and publicly justifiable. An EU court that substantially relies on an AI risk assessment score in a sentencing decision without articulating independent reasoning violates Art.6(1) ECHR right to a reasoned judgment.

ODR Regulation 2023/2440 + EU AI Act: When Alternative Dispute Resolution Becomes Point 8

The EU's Online Dispute Resolution (ODR) framework creates a specific intersection with Annex III Point 8 through the revised ODR landscape and the EU AI Act's "similar way in alternative dispute resolution" language.

EU ODR Platform (Regulation (EU) No 524/2013): The European Commission's ODR Platform handles consumer-to-business disputes in e-commerce. The platform routes complaints to certified ADR bodies in each Member State. AI-assisted triage, evidence analysis, and outcome recommendation within the ODR platform process may trigger Point 8 when AI is used by the ADR body (not the consumer) to assess the merits of the dispute.

Revised ODR Landscape (2023/2440): Commission Implementing Regulation (EU) 2023/2440 on the ODR platform's operational rules includes provisions for AI-assisted dispute analysis. When AI generates a binding recommendation or a non-binding settlement proposal that is determinative in practice (because parties accept it rather than escalating to formal litigation), the AI is functioning "in a similar way in alternative dispute resolution" — triggering Point 8 scope.

Smart Contract Arbitration AI: Arbitration AI embedded in smart contract dispute resolution platforms (e.g., Kleros, Aragon Court) presents a novel Point 8 question. If these platforms are used by EU-registered arbitration centres or endorsed by EU dispute resolution bodies, the AI components that determine outcomes may qualify as Point 8 high-risk — even when running on blockchain infrastructure.

Art.6 ECHR Fair Trial + Art.14(4) EU AI Act — The Dual Constitutional Constraint

The constitutional framework for judicial AI in the EU is defined by two parallel obligations that together prohibit fully automated judicial decisions.

Art.14(4) EU AI Act — No Automated Individual Decisions in High-Risk Contexts: For all Annex III high-risk AI systems deployed by public authorities, Art.14(4) mandates that human oversight measures ensure that no decision affecting a natural person is taken solely on the basis of the AI system's output. For Annex III Point 8 AI specifically, this means: a judge must review AI-generated legal analysis, fact interpretation, or law application independently before issuing a judgment or order. A judicial decision that simply adopts an AI output without independent judicial reasoning violates Art.14(4).

Art.6 ECHR Fair Trial — Individual Reasoning and Judicial Independence: The European Court of Human Rights has established (Taxquet v. Belgium, 2010; Salabiaku v. France, 1988) that Art.6(1) requires judicial decisions to be supported by sufficient reasoning that allows the parties and the public to understand the basis of the decision. An AI-generated legal analysis that the judge adopts wholesale — without articulating why the AI's reasoning is correct and applicable — fails the Art.6(1) reasoning requirement even if the AI analysis is technically accurate.

The Combined Constraint for Deployers (Courts):

VwVfG §35a in German Administrative Judicial Context: German Administrative Procedure Act §35a prohibits fully automated individual administrative decisions unless statutory law expressly authorises them. For German administrative courts reviewing AI-generated administrative decisions (e.g., tax assessment AI, benefit eligibility AI), the court itself must conduct independent merits review — it cannot simply validate the administrative AI's output. This creates an additional layer: German administrative court AI tools used to review AI-generated administrative decisions are subject to Point 8 obligations.

German §261 StPO — AI-Assisted Evidence Evaluation in Criminal Proceedings

The German Code of Criminal Procedure (Strafprozessordnung — StPO) creates specific constraints on how AI can be used in German criminal proceedings. These constraints operate in addition to EU AI Act Point 8 obligations for any AI deployed in German criminal courts.

§261 StPO — Grundsatz der freien richterlichen Beweiswürdigung (Free Judicial Evaluation of Evidence): The fundamental principle of German criminal procedure is that the trial judge must freely evaluate all evidence presented, forming their own conviction (Überzeugung) as to the facts. §261 StPO explicitly requires the court to form its conviction from the entirety of the trial proceedings. This means: AI-generated evidence analysis is admissible as expert input, but the judge cannot outsource the conviction-forming process to the AI. An AI that tells the court "the defendant is guilty with 87% probability" violates §261 StPO if the judge treats that probability as determinative.

§244(3) StPO — Evidence Admissibility Threshold: Evidence that is insufficiently reliable may be excluded under §244(3) StPO. AI-generated evidence analysis (e.g., deepfake detection, linguistic authorship attribution, forensic AI analysis) must meet reliability standards established by German courts. The BGH has held that courts must critically assess the reliability of scientific methods before admitting expert opinions based on them — this applies equally to AI-based forensic analysis.

§§ 72-93 StPO — Expert Evidence Framework: AI systems used as the basis for court-appointed expert opinions (e.g., forensic voice analysis AI, gait analysis AI, document forensics AI) must comply with the expert evidence framework. The expert (human) bears responsibility for the opinion under §78 StPO — AI is a tool, not a co-expert. If an expert cannot explain how the AI reached its conclusion, the expert opinion may not meet the explainability standards required under §261 StPO (court must be able to evaluate the basis of conviction).

BVerfG 2 BvR 2628/10 — Right to Be Heard in Algorithmic Decisions: The Federal Constitutional Court has established that where a court relies on computer-generated or algorithmic analysis in a decision, the parties must be given adequate opportunity to challenge the reliability and methodology of that analysis. For EU AI Act Point 8 deployers (German courts): defendants and parties in proceedings must be informed when AI tools have been used in judicial analysis (Art.50 EU AI Act transparency obligation), and must be given the opportunity to challenge the AI's output under Art.26(5) human oversight provisions.

Eurojust as Institutional Actor — AI in Judicial Cooperation

Eurojust (European Union Agency for Criminal Justice Cooperation, The Hague) is the EU's judicial cooperation body, distinct from Europol (law enforcement) and OLAF (anti-fraud). Eurojust is subject to the EU AI Act as an EU institution, body, office, or agency under Art.2(1)(b).

Eurojust AI Systems in Active Use:

Eurojust as Provider vs. Deployer: When Eurojust develops AI systems that it provides to national judicial authorities for use in cross-border criminal proceedings, Eurojust may function as a provider under Art.3(3) EU AI Act. National judicial authorities deploying Eurojust-provided AI for case-specific legal analysis then bear deployer obligations under Art.26. The provider/deployer split for EU agency-to-national-authority AI distribution creates compliance complexity: which authority registers in the EU AI Act database under Art.71? Which authority conducts the conformity assessment? No published guidance from Eurojust addresses this as of 2026.

Distinction from Europol (Point 6): Europol AI systems are classified under Annex III Point 6 (law enforcement) when they assist police operations. Eurojust AI systems may qualify under Point 8 when they assist the judicial phase (prosecution and court stage) of proceedings. The institutional divide between police intelligence (Point 6) and judicial cooperation (Point 8) is operationally significant for EU AI Act compliance planning.

Python JudicialAIClassifier

from dataclasses import dataclass
from typing import Literal

@dataclass
class JudicialAISystem:
    name: str
    primary_function: str
    competent_authority_context: bool
    deployment_context: Literal["court", "adr_authority", "law_firm", "private_party", "court_expert"]
    decision_type: Literal["fact_research", "law_research", "law_application", "risk_scoring", "procedural", "aggregate"]
    provider_jurisdiction: Literal["eu", "us", "uk", "other"]
    assists_final_decision: bool

def classify_judicial_ai(system: JudicialAISystem) -> dict:
    # Prohibited — fully automated judicial decisions
    if system.assists_final_decision and system.decision_type in ["law_application", "risk_scoring"]:
        if system.competent_authority_context:
            return {
                'classification': 'PROHIBITED_USE',
                'trigger': 'Art.14(4) — no automated individual judicial decisions',
                'notes': 'Deployment design must include mandatory human review before any output affects proceedings'
            }

    # Point 8 high-risk triggers
    if system.competent_authority_context and system.deployment_context in ["court", "adr_authority", "court_expert"]:
        if system.decision_type in ["fact_research", "law_research", "law_application", "risk_scoring"]:
            cloud_flag = system.provider_jurisdiction in ["us", "uk"]
            return {
                'classification': 'HIGH_RISK',
                'trigger': 'Annex III Point 8(a)',
                'cloud_act_flag': cloud_flag,
                'notes': 'Full Title III obligations. Art.14(4) human oversight mandatory. Art.50 transparency to parties.'
            }

    # Not high-risk — law firm or private party context
    if system.deployment_context in ["law_firm", "private_party"]:
        return {
            'classification': 'NOT_HIGH_RISK',
            'trigger': 'No competent authority context',
            'notes': 'Standard EU AI Act obligations apply (transparency if GPAI model used)'
        }

    # Procedural/aggregate — not Point 8
    if system.decision_type in ["procedural", "aggregate"]:
        return {
            'classification': 'NOT_HIGH_RISK',
            'trigger': 'Procedural or aggregate — no substantive legal analysis',
            'notes': 'Ensure system does not drift into substantive fact/law analysis over time'
        }

    return {
        'classification': 'AMBIGUOUS',
        'trigger': 'Manual legal opinion required',
        'notes': 'Edge case — obtain specific EU AI Act Art.6 classification analysis'
    }

# Test classifications
test_systems = [
    JudicialAISystem("Harvey AI for EU Court", "Legal research for judge", True, "court", "law_research", "us", False),
    JudicialAISystem("Harvey AI for Law Firm", "Legal research for lawyer", False, "law_firm", "law_research", "us", False),
    JudicialAISystem("Luminance EU Arbitration", "Document analysis for arbitrator", True, "adr_authority", "fact_research", "uk", False),
    JudicialAISystem("COMPAS-EU Recidivism", "Recidivism risk score for sentencing", True, "court", "risk_scoring", "us", False),
    JudicialAISystem("Sentencing AI Auto-Decision", "Automated sentence generation", True, "court", "law_application", "eu", True),
    JudicialAISystem("ODR Settlement AI", "Settlement recommendation for ODR arbitrator", True, "adr_authority", "law_application", "eu", False),
    JudicialAISystem("Court Scheduling AI", "Hearing scheduling and case routing", True, "court", "procedural", "eu", False),
    JudicialAISystem("Eurojust Legal Analysis AI", "Comparative EU law for cooperation", True, "court_expert", "law_research", "eu", False),
    JudicialAISystem("Private ODR Prep AI", "Party preparing ADR submission", False, "private_party", "law_research", "us", False),
    JudicialAISystem("Court Analytics Dashboard", "Aggregate case outcome statistics", True, "court", "aggregate", "eu", False),
]

for s in test_systems:
    result = classify_judicial_ai(s)
    cloud = " [CLOUD ACT RISK]" if result.get('cloud_act_flag') else ""
    print(f"{s.name}: {result['classification']} ({result['trigger']}){cloud}")

Classifier output:

• Harvey AI for EU Court: HIGH_RISK (Annex III Point 8(a)) [CLOUD ACT RISK]
• Harvey AI for Law Firm: NOT_HIGH_RISK (No competent authority context)
• Luminance EU Arbitration: HIGH_RISK (Annex III Point 8(a)) [CLOUD ACT RISK]
• COMPAS-EU Recidivism: HIGH_RISK (Annex III Point 8(a)) [CLOUD ACT RISK]
• Sentencing AI Auto-Decision: PROHIBITED_USE (Art.14(4) — no automated individual judicial decisions)
• ODR Settlement AI: HIGH_RISK (Annex III Point 8(a))
• Court Scheduling AI: NOT_HIGH_RISK (Procedural or aggregate)
• Eurojust Legal Analysis AI: HIGH_RISK (Annex III Point 8(a))
• Private ODR Prep AI: NOT_HIGH_RISK (No competent authority context)
• Court Analytics Dashboard: NOT_HIGH_RISK (Procedural or aggregate)

The critical scope distinction: the same Harvey AI product is HIGH-RISK when licensed to a court and NOT_HIGH-RISK when licensed to a law firm. The provider's obligations activate when the system is "intended to be used" by a competent authority — foreseeability of court use in marketing or contracts establishes intent.

Annex III Point 8 Compliance Checklist (25 Items)

Provider Obligations (Art.16)

1. ☐ Determine whether your system qualifies as assisting a court/judge in fact research, law research, or law application — or functions similarly in ADR — before assuming no Point 8 obligation
2. ☐ Identify all actual and foreseeable deployment contexts — if courts, tribunals, or ADR authorities are known or likely users, Point 8 provider obligations apply regardless of primary market focus
3. ☐ Establish quality management system under Art.17 — document all training data sources and legal corpora used to train the system
4. ☐ Prepare technical documentation under Annex IV — include accuracy benchmarks on EU Member State legal corpora, not only US or UK case law
5. ☐ Implement data governance under Art.10 — ensure training data includes EU law, ECJ/ECHR case law, and Member State legal materials proportionate to intended deployment jurisdictions
6. ☐ Design mandatory human oversight measures under Art.14(4) — no judicial AI output may automatically become part of court record without judicial review; implement technical controls if possible
7. ☐ Conduct conformity assessment under Art.43 — third-party conformity assessment via notified body if no EU harmonised standard applies (no EN standard for judicial AI exists as of 2026)
8. ☐ Draft EU Declaration of Conformity and affix CE marking under Art.48-49 before placing system on EU market for court use
9. ☐ Register in EU AI Act public database under Art.71 before deploying in any EU judicial authority context
10. ☐ Implement post-market monitoring under Art.72 — collect accuracy, bias, and reliability data across legal systems, languages, and case types

Deployer Obligations — Courts and ADR Bodies (Art.26)

1. ☐ Verify provider has completed conformity assessment and CE marking before using any AI for case-specific fact or law analysis under Art.26(1)
2. ☐ Implement Fundamental Rights Impact Assessment (FRIA) under Art.26(10) — mandatory for all public authority deployers; assess specific impact on defendants, asylum seekers, and procedural fairness
3. ☐ Establish human oversight procedure under Art.26(5) — no AI-generated legal analysis may be incorporated in judicial decisions without articulated judicial review
4. ☐ Register as deployer in EU AI Act public database under Art.71(6)
5. ☐ Report serious incidents (AI error causing miscarriage of justice, procedural violation, or fundamental rights breach) to market surveillance authority under Art.73

Fundamental Rights and Transparency

1. ☐ Notify parties to proceedings when AI has been used in case analysis or preparation of judicial materials — Art.50 transparency obligation
2. ☐ Ensure parties can challenge AI-generated analysis — Art.26(5) requires mechanisms for parties to question AI outputs affecting their proceedings
3. ☐ Assess bias risk for protected characteristics under Art.9(9) — legal AI trained primarily on US/UK law may systematically underperform on EU Member State legal questions affecting defendants from non-Anglophone jurisdictions
4. ☐ Document legal basis for processing case file data under GDPR Art.6 and — where applicable — GDPR Art.9 (special category data in criminal proceedings)
5. ☐ Assess CLOUD Act exposure for any US-based or UK-based AI providers processing EU case files, draft judgments, or attorney-client communications

Art.6 ECHR and Constitutional Constraints

1. ☐ Ensure no AI-generated legal analysis or risk score is treated as sole or determinative basis for judicial decision — Art.6 ECHR individual reasoning obligation
2. ☐ Train judicial staff and ADR professionals on the distinction between AI-assisted research (permissible) and AI-determined outcomes (impermissible) — Art.26(4) deployer training obligation
3. ☐ Document judicial reasoning independently from AI output in all decisions where AI was consulted — internal case record should distinguish AI-assisted research from judicial analysis

German-Specific

1. ☐ Confirm AI-assisted evidence analysis used in German criminal proceedings meets §244(3) StPO reliability standards and that the human expert can explain AI methodology to satisfy §261 StPO judicial evaluation obligation
2. ☐ For German administrative courts: ensure AI tools used to review AI-generated administrative decisions comply with Point 8 obligations — the court's review of algorithmic administrative acts is itself a judicial-AI interaction requiring Art.14(4) oversight

EU AI Act Annex III High-Risk Categories — Series Navigation

This post is part of the EU AI Act Annex III High-Risk AI Categories series, covering all eight Annex III points for developers and compliance teams building AI systems subject to EU AI Act Title III.

See Also

• EU AI Act Annex III Point 6: Law Enforcement AI — recidivism and criminal risk AI overlaps Point 6 (police phase) and Point 8 (judicial phase); the same COMPAS-equivalent tool may trigger both points depending on deployment stage
• EU AI Act Annex III Point 5: Essential Services AI — Art.14(4) human oversight constraint for AI decisions affecting fundamental entitlements (benefits, credit) uses the same structure as Point 8 judicial decisions; the constitutional principle is identical
• EU AI Act Annex III Point 7: Migration and Border Management AI — Eurojust (Point 8) and Frontex (Point 7) AI systems share EU agency compliance architecture; no public conformity assessments published for either as of 2026
• EU AI Act Art.5: Prohibited AI Practices — Art.5(1)(b) social scoring prohibition applies before Point 8 conformity assessment; judicial risk AI that extends to generalised social trustworthiness scoring breaches Art.5(1)(b) rather than qualifying as compliant high-risk